Time Series Momentum Strategies Using Deep Neural Networks
Log in to collectAcademic paper
Enhancing Time Series Momentum Strategies Using Deep Neural Networks
Bryan Lim; Stefan Zohren; Stephen Roberts
- Quantitative BioSciences
- University of Oxford
- ?University of Oxford - Oxford-Man Institute of Quantitative Finance
Strategy in a nutshell
The strategy trades 88 futures across equities, commodities, FX, and fixed income, targeting 15% portfolio volatility while accounting for trading costs (~10 bps). An LSTM model estimates trends and position sizes simultaneously by optimizing the Sharpe ratio via a loss function incorporating returns adjusted for costs. Inputs include normalized returns (1-day to 1-year) and MACD indicators at multiple time scales. LSTM hidden and cell states, with input/output gates, process historical information effectively. The portfolio is rebalanced daily, using advanced trend estimation and position-sizing techniques for optimal multi-asset performance.
Economic rationale
Machine learning momentum models mimic traditional momentum by betting that winners keep winning and losers keep lagging. They enhance risk-adjusted returns by jointly optimizing trend estimation and position sizing. Calibrated using the Sharpe ratio, these strategies often outperform benchmarks, though frequent rebalancing incurs trading costs. Using liquid contracts or incorporating costs into model training mitigates this, enabling higher net returns than traditional approaches.